Andrei Alexandru

The catalyst for a change in how we do our business.

AI Engineer · London, UK

i

Praise

What they said

What consistently stood out about Andrei was his combination of autonomy and collaboration... His problem-solving approach is structured, clear, and grounded in practical engineering judgment.
Andrew Solomon, Lead Generative AI Engineer, Jaguar Land Rover

What sets Andrei apart is his proactive mindset and intellectual curiosity. He does not wait to be told what to do; he identifies opportunities and acts on them.
Huseyn Gorbani, AI Developer, RLB UK

ii

Chapter One

Practice

Words first, then the system.

I write the thing before I build it. Always have. For Renda that meant a spec long enough to bore anyone who opened it, finished before a line of code existed. The spec is the engineering. What a system does, where it breaks, how you would know: that part is not typing, and it does not get faster.

Agents implement against it, test first. Review, architecture and verification stay with me. That is how one person ships and maintains a product this size while holding down a full time job.

It is also why I built No-Numb. Build fast with an agent and you can end up not understanding your own codebase. That is a bad feeling and it creeps up on you. So I wrote a tool that quizzes me on the code before it lets me carry on. It turns out I am not the only one who worried about that.

The habit runs past code. The assistant on my server keeps its behaviour in Markdown, versioned in git. When I want it to act differently I do not touch the code. I change the words and deploy them. At Jaguar Land Rover I read the entire EU AI Act. All of it. Then I wrote what we were supposed to do about it.

The day job is production LLM systems. Pipelines on AWS Bedrock, orchestrated with Step Functions. Structured output with schema enforcement. Model routing that escalates by risk. Document extraction that never lets private data linger. Before that, RAG applications on Azure, and Computer Vision research in automotive.

1

Chapter Two

Work

Automotive research, then enterprise AI, with a company of my own running alongside it since 2025.

CurrentApr 2026 to nowRemote

AI Engineer

Citation Group

Building production LLM pipelines on AWS Bedrock, orchestrated with Step Functions
Serverless Python services on Lambda, provisioned end to end with Terraform
Structured-output LLM workflows with risk-based model routing and tiered escalation
Document extraction and OCR pipelines over S3 and DynamoDB, built PII-safe by design

CurrentSep 2025 to nowLondon

Co-founder / Engineering Lead

Alimio Ltd

Co-founded a two-person UK company and built the technical side of both products
Renda: SaaS that turns Claude Design exports into decks, carousels and PDFs, past 1000 users on organic acquisition alone
Alimio: AI food waste app with photo scanning and voice input, published on the App Store and Google Play
Own the full stack for both: architecture, edge infrastructure, LLM layer, billing and deployment

Mar 2025 to Apr 2026London

AI Developer

RLB (Rider Levett Bucknall)

Built RAG-based POC applications for enterprise clients
Deployed and fine-tuned LLMs on Azure
Full Azure AI stack (OpenAI, AI Foundry, AI Vision, AI Speech)

Sep 2023 to Mar 2025Coventry

Graduate AI Engineer

Jaguar Land Rover

Computer Vision research in the AI department
Built and evaluated LLM applications for internal tooling
Contributed to AI Governance frameworks and EU AI Act compliance

Jun 2021 to Aug 2022Coventry

Software Intern (Industrial Placement)

Jaguar Land Rover

Used ML to optimize Jaguar I-PACE traction control system
Wrote C code for traction control algorithms
Contract extended by 2 months due to performance

2

Chapter Three

Built

Shipped, with users. Not prototypes.

Fig. 3.12026

People were screenshotting their Claude Design work and posting it blurry. That was the whole problem. Renda takes the export, standalone HTML or a project ZIP, and gives back clean PNGs, LinkedIn carousels, PDFs and PowerPoint decks. I built the technical side. My co-founder found the people. It passed a thousand users without us paying for a single one.TypeScript, Cloudflare Workers, Hono, React 19, Vite, Google Vertex AI, Supabase, Polar

Fig. 3.22025

Everyone throws food away. We wanted to know whether a phone camera could stop it. Point it at a packet and Alimio reads the product and the date off the label. Talk to it and it writes the list down. When something is about to go off it suggests what to cook with it. On the App Store and Google Play.Flutter, Azure AI Vision, Azure AI Speech, GPT-4, Azure AI Foundry

Fig. 3.32025 - 2026

Nuree AI — A founder came to me with an idea and no engineer. I was the whole technical side. We built a voice coach with a 3D avatar that talks you back to work when your attention goes, and it had to interrupt and be interrupted the way a person does. It ran until March 2026.Next.js, Vercel, Supabase, ElevenLabs, OpenAI, 3D Animation

Fig. 3.42026

I noticed I was shipping code I could not explain. No-Numb waits until the agent has edited a file, then makes it quiz me on what it just did. Get one wrong and it shows the answer and why, then asks again. The session does not move on until I pass. The friction is the point.Bash, jq, Claude Code Hooks, Plugin Marketplace, Markdown Skills, JSON

Proactive AI Assistant

I did not want another assistant I had to ask. I wanted one that knew when to speak. It runs on a private server, reads my mail, my calendar and a summary of my spending, and wakes on a schedule I cannot predict. A cheap model decides whether anything matters. A better one writes the message, or throws it away. Most of the time it says nothing, which is the part I am proudest of.

Open Banking Finance Platform

Four bank accounts and no real idea what was happening across them. This pulls all four through Open Banking twice a day, sorts every transaction with a model, and finds the subscriptions I forgot about and the charges that look wrong. It hands a summary to the assistant above through one authenticated endpoint, so that thing never gets near my bank credentials.

3

Appendix

Ask the AI

An AI assistant, not Andrei. It has read this book and nothing else, and runs on Gemini 3.7 Flash. For anything it cannot answer, email him.

Q. Where does Andrei work now?

He is an AI Engineer at the Citation Group, remote from London, since April 2026, building production LLM pipelines on AWS Bedrock orchestrated with Step Functions. Alongside that he co-founded Alimio Ltd, where he builds the technical side of Renda and Alimio.

4

Note

Origins

BSc Computer Science, First Class Honours, University of Surrey.

I wanted to know why a neural network decided what it decided. So I built a way to pull a readable decision tree out of a trained one, and tested it on image classification and on bank fraud.

Reduced decision tree complexity by up to 70%
Held accuracy within a few percentage points of the original networks
Relevant to XAI, model interpretability and EU AI Act compliance

At the time it was a curiosity. Interpretability is written into law now, and reading the EU AI Act at Jaguar Land Rover felt like running into my own dissertation from the other side. The code is on GitHub.

5

Index

What I work with

Languages: Python, TypeScript, JavaScript, C, SQL
AI & ML: LLMs & Prompt Engineering, RAG Systems, LLM Orchestration, Agentic Systems, Structured Output, Computer Vision, Model Fine-tuning, MLOps
Cloud & infrastructure: AWS Bedrock, AWS Lambda, Step Functions, Cloudflare Workers, Google Vertex AI, R2 & KV, Terraform, Azure OpenAI, Azure AI Foundry
Frontend: Next.js, React 19, Vite, Flutter, Tailwind CSS, Framer Motion
Tools: Git, GitHub Actions, Claude Code, pytest, Vitest, Hono, Supabase, Polar

Andrei Alexandru, AI Engineer, London, UK.

alexandruciprian61@gmail.com

LinkedIn · GitHub · CV

Set in Literata and IBM Plex Mono.

6